fix(rooms): carry cross-segment turn bookends in historyPolicy: full

[lsaether]

Context

After #58 shipped rooms with mid-turn rotation deliberately preserved (an active turn is not torn down across segment boundaries), historyPolicy: full could leave a late joiner with an orphan: amux/turn_complete lands in the active segment, but amux/turn_started for the same turn lives in the closed prior segment and was filtered out by the current-segment-only rule. Coterie- and vibe-style correlators either drop the orphan or render the finished turn as if the agent were still typing.

Scoped subset of #57 (the broader "orphan chunks outside any turn" case remains open).

Fix

history_full and replay_entries_for_policy now compute a cross_segment_turn_carry set:

• Every amux/turn_started / amux/turn_complete / amux/turn_cancelled frame found in the active segment contributes its amuxTurnId.
• The currently active turn's amuxTurnId (from active_amux_turn_id) is also added so an in-flight cross-segment turn's turn_started gets carried even before it completes.

Frames that are turn-lifecycle frames with an amuxTurnId in the carry set are then included in the full view even when they sit in a prior segment. Non-lifecycle frames (chunks) from prior segments are still excluded — those belong to full_lineage.

Frames stay in original replaySeq order, so the carried turn_started appears temporally before the active-segment frames and the bracket reads naturally on the client.

Wire impact

None — strictly additive. The full view returns more frames it should have been returning all along. No new fields, no new methods, no enum variants. Today's clients just stop seeing the orphan they couldn't handle.

Tests

Four new unit tests in src/room/state.rs::tests:

1. history_full_carries_turn_started_when_turn_completed_across_segments — turn_started in seg-1, hermes compaction mid-turn, turn_complete in seg-2; full view contains both bookends in order.
2. history_full_carries_turn_started_when_turn_is_still_active_across_segments — turn_started in seg-1, rotation, turn still in flight; carry pulls turn_started forward via active_amux_turn_id.
3. history_full_does_not_carry_turn_that_completed_before_active_segment — a fully prior-segment turn must NOT appear in full view.
4. history_full_lineage_returns_every_segment_unchanged — regression guard that lineage view is unaffected.

cargo fmt --check, cargo clippy --all-targets -- -D warnings, 63 lib + 78 integration tests all pass.

Test plan

• cargo fmt --check
• cargo clippy --all-targets -- -D warnings
• cargo test (63 lib + 78 integration)
• Manual smoke against real hermes that compacts mid-turn — deferred; covered by mock-based tests since hermes-acp 0.14 doesn't expose _meta.hermes yet anyway.

[lsaether]

Pushed fe167c8 on top — additive, no force.

Docs

• docs/design/rooms.md ## historyPolicy — full description now spells out the lifecycle-frame carve-out (amux/turn_started / amux/turn_complete / amux/turn_cancelled from prior segments carried when bracketing the active segment) and the reason (mid-rotation turn stays bracketed), plus the guarantee that non-lifecycle frames stay excluded.
• src/protocol/attach.rs HistoryPolicy::Full rustdoc — same point in shorter form for the rustdoc surface.
• README.md — replay-log bullet and session/attach coverage row both call out the carve-out so the README doesn't drift from rooms.md.

CHANGELOG

• New Fixed entry under Unreleased pointing at the cross-segment turn bookend fix as a scoped subset of Normalize orphan replay chunks during backfill #57.

Marking the PR ready for review.